Dpis Runtime Route Diagnose

DPIS Runtime Route Diagnose

This is a project-local skill bundle. It is intentionally stored in the repository instead of a global Codex skill, so new conversations can discover DPIS runtime rules through AGENTS.md without modifying user-wide skill files.

Trigger

Use this skill before diagnosing or changing behavior involving:

• LSPosed / Xposed runtime hooks
• app flicker, flash, relaunch, or configuration churn
• missing DPI, viewport, smallest-width, or font scaling
• font hook chains, route domains, or template recommendations
• system_server, ActivityThread, Resources, Display, WindowMetrics, WebView, TextView, Paint, Flutter, or HyperOS font routes
• shared runtime code under app/src/main/java/com/dpis/module/

Do Not Use

Do not use this skill for:

• pure UI layout work
• release-note wording
• ordinary Java refactors that do not affect runtime hook behavior
• CI failures unless the failure changes runtime route behavior

For Java / CI API compatibility questions, read docs/java-toolchain-policy.md instead.

Required Reads

Read only the documents needed for the touched route:

• Always read AGENTS.md.
• For font routes, read docs/font-routing.md.
• For Modern routes, read docs/modern-runtime-resync.md.
• For Legacy routes or shared code that may affect Legacy, read docs/legacy-runtime-resync.md.
• For log collection, read docs/lsposed-diagnostics.md.
• For Java / CI API compatibility questions, read docs/java-toolchain-policy.md.

Workflow

1. Classify the layer that owns the suspected behavior.
2. Gather the smallest evidence set that can prove or disprove that layer.
3. Probe route execution one boundary at a time.
4. Choose the narrowest fix shape that expresses the route semantics.
5. Add or update focused tests for policy behavior.
6. Update the relevant living route document.

Semantic Frame

Before editing, name the owning layer:

• UI preference or form draft
• recommendation template
• planner domain
• app-process route
• system_server route
• lifecycle entry
• mutation field
• config source
• LSPosed scope or module loading

If two layers appear to write the same state, resolve the ownership boundary before adding another branch.

Probe Order

Probe one boundary at a time:

1. Module entry reached.
2. Guard did not return early.
3. Dependency or class lookup succeeded.
4. Hook installed.
5. Callback fired.
6. Target package resolved.
7. Config source returned the expected app config.
8. Field policy allowed the mutation.
9. Mutation changed the runtime object.
10. App-visible behavior changed.

hook ready proves only that installation ran. Require a callback, mutation, counter, or visible result before calling a route effective.

Diagnostic Evidence Rules

When reviewing an exported diagnostic package:

• Separate the diagnostic session window from the Perfetto trace window. Do not generalize a short trace to the whole session.
• Treat measuredCalls=0, calls=0 with non-zero applied/skipped, and zero latency percentiles as missing or inconsistent instrumentation, not as proof of zero overhead.
• Treat stable_metrics, stable_configuration, and no_viewport_delta_after_resolution as policy skips. They are not failures unless the selected config still required a change.
• Treat an empty diagnostic marker as incomplete evidence. Cross-check raw lsposed-log.txt and runtime-hotpath events before declaring a route absent.
• A module-effects.tsv row selected but no … route effect observed is valid only when the session window also has no matching structured hot-path event. WeChat DPI mutations are route=wechat_dpi runtime events, not ProcessPerformance aggregate route names. Font aggregates plus a WeChat mutation_applied event must not be read as “WeChat DPI did nothing”.
• Prefer the last stage=aggregate snapshot per process. Earlier snapshots are cumulative prefixes, not a second independent session.
• Font rewrite applied details should include in=/out=. Missing pairs on textview_current_px_fallback are an evidence-format gap, not proof the route skipped. Timeline applied counts can be lower than the aggregate because kept is aggregate-only and transport may sample begin/applied/end.
• Do not treat |incoming - earlier out| < epsilon as double-scale. A second-order candidate is an incoming px that equals a first-scale output. Independent later sizes, including another object's unscaled 28.2, must still rewrite. Same-object already-scaled Paint must not.
• A hot-reload warning is recoverable only if later module-entry and route evidence exists after the restart/reload path. Otherwise report it as an unresolved framework or loading failure.
• For performance claims, check thread_state and sched_slice for the target process and main thread. Distinguish Running, Runnable, and sleeping or blocked states before naming a bottleneck.
• For exported ZIP or extracted diagnostic directories, run python tools/analyze_diagnostic_pack.py <zip-or-dir> before narrating font/DPI effectiveness. Use --self-check to verify the analyzer itself.